DuckDB

Why it matters

Analytics used to mean standing up infrastructure. DuckDB recognized that most analytical work is one person, one laptop, files on disk — and built a vectorized query engine for exactly that. It reads Parquet directly, parallelizes aggressively, and embeds anywhere C++ or Python runs.

Engineering notes

• In-process, zero-copy reads. Queries run inside your application; Parquet files are queried in place without loading.
• Vectorized execution. Processes data in batches sized for CPU caches — the same design principle as the big columnar warehouses, without the cluster.
• A real SQL dialect. Window functions, CTEs, structs, and sensible type coercion — not a toy subset.

Limitations

• Not a replacement for a shared warehouse when many writers or concurrent users are involved.
• The extension ecosystem is younger than Postgres'; check for your niche format before committing.